Skip to content

Releases: neeveew/AI-Arena-Adversarial-LLM-Lab

AI Arena Lite 0.4.141-beta

Pre-release

Choose a tag to compare

@neeveew neeveew released this 09 Sep 06:19

AI Arena Lite is the C# Windows edition of AI Arena. This beta adds live transcript streaming, clearer model controls and improved context handling.

Highlights

  • Stream Arena participant and Narrator answers into one live transcript card, then commit the same card when the turn completes. Preserve reader position and keep canceled or interrupted previews separate from saved history.
  • Show provider-observed Loading model, Reading context, Thinking, and Writing stages, with a clear waiting state when the server does not report its activity.
  • Budget requests using the loaded model's current native server context allocation instead of a catalog maximum. Preserve Strict history; let Rolling 80% omit older whole entries while retaining required conversation text.
  • Recover completed reasoning-only responses with at most one supported off/low reasoning retry on the same model and exact conversation. Send LM Studio reasoning off explicitly and preserve the original native conversation anchor.
  • Distinguish empty public answers from provider rejection. Preserve reported stop codes and token counts, keep output-cap inference separate, and never treat partial public output, incomplete streams, or tool output as reasoning-only recovery candidates.
  • Keep Factory retries tied to their original retained public entries and causal fingerprint. Show the opening Public Operator composer in the empty transcript with Send only and Send and start Auto Chat actions.
  • Expand each selected model's compact settings beneath its catalog row. Keep context length and agent assignment immediately visible, give expanded content a clear boundary, and prevent competing nested scrolling.
  • Update the offline Help Center for model configuration, streamed responses, context fitting, and reasoning-only recovery. Continue saving configuration automatically.

Assets

  • AI.Arena.Setup.0.4.141-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • installer-signing.json
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

Installation details

Windows x64, self-contained with the .NET Desktop Runtime, offline Help and optional local SearXNG support. This beta is unsigned; signing reports and SHA-256 checksums accompany the installer. The separate C++ app remains in development.

Installer checksum

AI.Arena.Setup.0.4.141-beta.exe (106,010,441 bytes):

5c1d59c677f0f66ac957d757e28caee1d4e252db2a2f2640a00134197dc35b3f

GitHub replaces spaces with dots in the download filename. The unchanged SHA256SUMS.txt records the original filename, AI Arena Setup 0.4.141-beta.exe; both names refer to the same verified bytes.

AI Arena Lite 0.4.140-beta

Pre-release

Choose a tag to compare

@neeveew neeveew released this 08 Sep 11:27

AI Arena Lite is the C# Windows edition of AI Arena. This beta brings together the improvements since the last public release, 0.4.134-beta.

The native C++ app remains in development and is not included in this installer.

Highlights

  • Discover running LM Studio, Ollama, llama.cpp, and compatible model servers automatically, and choose supported communication and model-management APIs from server evidence.
  • Use models from multiple servers in one arena. Keep assignments and model settings tied to their serving endpoint, including models with identical names on different servers.
  • Simplify Provider connection to server status, Find servers, Manage models, and one Advanced section. Remember the last session and configuration without requiring saved connection profiles.
  • Preserve explicit model routes and unfinished connection drafts. Reject stale discovery, catalog, assignment, and credential results so newer settings and cleared tokens stay authoritative.
  • Add optional Native Services communication with a separately installed, compatible AI Arena C++ app for diagnostics, model operations, and bounded inference. Keep cancellation, uncertain submissions, and operation receipts explicit; each app retains its own data profile.
  • Carry forward the Lite identity, green emblem, refreshed Help, and distinct shortcut icon. Install in Program Files\AI Arena Lite, safely migrate only the verified legacy per-user installation, and retain local user data and the sibling AI Arena shortcut.
  • Make the Universal Status Center the shared owner of Arena readiness warnings and blockers, with prerequisites available from disabled controls.
  • Preserve reader position and selection in Agent and AI Collaborate, count unseen messages, and provide Jump to latest. Keep live response state stable during theme refresh.
  • Make AI Collaborate document import asynchronous, cancellable, and bounded, with supported text encodings, duplicate handling, and partial-success receipts.
  • Validate numeric Advanced settings before saving, show accessible field errors, and focus the first invalid field.
  • Protect and restore unsent Operator, Agent, and AI Collaborate drafts. Improve conversation persistence and retain drafts after save failures, cancellation, or in-flight edits.
  • Add bounded Trash with confirmation and Undo for sessions and restore points. Create safety checkpoints before replacing live Arena state, and report committed saves accurately when later logging or refresh fails.
  • Share session loading and snapshot-change detection across startup, navigation, refresh, and search. Cancel superseded cross-session searches and stop narration when the active session changes.
  • Consolidate privacy-safe workflow and provider errors with stable support codes. Retain accepted download receipts through observation failures and report uncertain model lifecycle outcomes accurately.
  • Restrict automatic completion retries to eligible transient rejections with an explicit end-to-end idempotency contract. Reuse the same request and key, and never replay accepted, partial, ambiguous, timed-out, or canceled completions automatically.
  • Package the self-contained Windows x64 app, required .NET Desktop Runtime, offline Help, and optional local SearXNG search component. Release metadata, dependency inventory, signing reports, and checksums accompany the installer.

Assets

  • AI.Arena.Setup.0.4.140-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • installer-signing.json
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

Installation details

Windows x64, self-contained with the .NET Desktop Runtime, offline Help and bundled local SearXNG support. Installs into Program Files\AI Arena Lite and preserves user data during migration. This beta is unsigned; signing reports and SHA-256 checksums are included with the release.

Installer checksum

AI.Arena.Setup.0.4.140-beta.exe (105,903,336 bytes):

e19fa0eab3da97900164f88ee97ac597ca8e9367235c50387e37d8a732de62da

GitHub replaces the installer's spaces with dots in its download filename.
The unchanged build checksum file, SHA256SUMS.txt, records its original name:
AI Arena Setup 0.4.140-beta.exe. Both names refer to the same verified bytes.

AI Arena 0.4.134-beta

AI Arena 0.4.134-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 14 Aug 17:56

AI Arena 0.4.134-beta

Highlights

  • Security: update the bundled SearXNG HTTP/2 dependency from h2 4.3.0 to 4.4.1, which rejects duplicate Host request headers addressed by CVE-2026-71554 / GHSA-6hr6-w5qg-qmwg, while retaining an official-PyPI SHA-256 lock and payload inventory attestation.
  • Add Experiment Lab and Verification Lab with resumable experiment matrices, historical forks, versioned scenarios and benchmarks, blind or provider-backed judging, claim provenance, deterministic fault injection, black-box control commands, and evidence-sealed visual and runtime QA.
  • Add first-class right-rail Model Comparison & QA with model-neutral baselines, exact replay evidence, bounded local history, and explicit ready, partial, and blocked gates.
  • Add first-class user-owned llama.cpp support with OpenAI-compatible chat, capability-detected health and router controls, bounded retry behavior, truthful unavailable telemetry, and no inferred RAM or VRAM claims.
  • Expand Agent with Roslyn Impact Explorer, conservative project and package root-cause findings, approval-gated Solution Doctor repair diffs, focused verification, and commit or release-baseline comparison.
  • Modernize the WPF visual system with semantic tokens, responsive layouts, accessible state cards and popups, compact transcript controls, stronger keyboard and UI Automation behavior, and Dark, Light, High Contrast, reduced-motion, and scaling coverage.
  • Rebuild Models & assignments around pinned loaded models, a virtualized available catalog, provider-backed residency controls, dynamic Default and explicit routing, exclusive assignment transfers, bounded provider discovery, and durable per-model configuration.
  • Add Factory public-group memory, causal turn handling, reliable Auto Chat and 1 Turn readiness, explicit all-agent routing prerequisites, and privacy-safe evaluation evidence.
  • Add per-model context windows, Strict and Rolling 80% history, response tone, typed context recovery actions, and a compact universal Status Center with causal history and responsive access.
  • Rebuild the offline User Guide as a modern searchable Help Center with 20 versioned articles, guided journeys, contextual routes, rich Markdown, responsive navigation, focus restoration, and deterministic content-freshness gates.
  • Improve busy-session performance with virtualized conversations, incremental projections, prepared turn context, structural source caching, split telemetry cadences, cached hardware presentation, deadline-driven status scheduling, indexed Help search, and DNS validation reuse.
  • Strengthen local durability and release production with canonical experiment serialization, bounded artifact loading, batched recoverable event logging, privacy-safe snapshots, fail-closed Authenticode readiness, isolated parallel harnesses, authenticated receipts, deterministic checksums, and installer-resume binding.
  • Harden post-0.4.133 reliability with portable streamed SHA-256 checks in Solution Doctor, coherent transcript search across rapid external rewrites, hosted-runner-safe layout fixtures, stable noisy-command-capture test coverage, and exit-clean expected-failure release tests.

Assets

  • AI Arena Setup 0.4.134-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • installer-signing.json
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.127-beta

AI Arena 0.4.127-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 29 Jul 23:56

AI Arena 0.4.127-beta

Highlights

  • Adds Agent's native .NET Solution Doctor with bounded solution/project discovery, typed restore and build diagnostics, project-correct Build/Test/Run plans, safe command previews, and structured repair evidence.
  • Hardens Solution Doctor across project-only, partial, empty, equivalent, duplicate-path, and hostile-path workspaces while keeping restores separately approved and refusing ambiguous repair guesses.
  • Keeps the AI Lab top rail aligned and stable, preserves Match Setup as a same-position toggle, and moves actionable or transient helper status into a compact bottom-left dock that cannot change top-rail height.
  • Redesigns transcript cards with compact model statistics and detailed hover/focus telemetry, full-width collapsible reasoning, and permanently visible quiet 30-DIP Copy, Speak, Pin, Retry, Delete, and Compare controls.
  • Preserves responsive, keyboard, automation, reduced-motion, Dark Blue, Light, and High Contrast behavior with expanded regression coverage for shell layout, transcript cards, and Solution Doctor workflows.

Assets

  • AI Arena Setup 0.4.127-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • installer-signing.json
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.124-beta

AI Arena 0.4.124-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 29 Jul 01:26
ce96784

AI Arena 0.4.124-beta

Highlights

  • Reports a skipped arena operation as skipped rather than completed. Asking the PowerShell control plane for five concurrent turns reported five successes when three turns actually ran, and the run-loop event fired for the two that never happened. The button is disabled while the arena is busy, so nobody clicking could reach this, but every automated caller was told its turn had completed whether or not it had. Turn and narration commands now answer not_available when the arena was already busy, and the completion event is gated on the same answer.
  • Keeps the WPF PowerShell plane enabled by default, retains the self-screenshot command and bundled SearXNG Internet backend, and remains a self-contained Windows installation with no separate .NET Desktop Runtime requirement.

Assets

  • AI Arena Setup 0.4.124-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.123-beta

AI Arena 0.4.123-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 28 Jul 20:36

AI Arena 0.4.123-beta

Highlights

  • Refuses to type into a field that is disabled or read-only when the PowerShell control plane is asked to fill one. Setting the text worked regardless of either state, so a script could fill a box that the app had deliberately disabled while an operation was running, and be told it succeeded.
  • Keeps every crash report when an app failure cascades. Reports are named to the millisecond, and a failing app rarely fails once, so several failures inside the same millisecond previously overwrote each other and left only the last. The first report is usually the one worth reading.
  • Keeps the WPF PowerShell plane enabled by default, retains the self-screenshot command and bundled SearXNG Internet backend, and remains a self-contained Windows installation with no separate .NET Desktop Runtime requirement.

Assets

  • AI Arena Setup 0.4.123-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.122-beta

AI Arena 0.4.122-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 28 Jul 20:34
a4eecb2

AI Arena 0.4.122-beta

Highlights

  • Restores the right rail after the window is made narrow and widened again. Crossing the narrow threshold once hid it for good: the saved preference still said the rail should be shown, nothing announced the change, and only the F7 shortcut brought it back.
  • Stops App Settings opening on top of Match Setup. Settings covered Match Setup's close button and primary actions, so the only way out was to close Settings first. Opening Match Setup has always hidden Settings; the reverse now matches.
  • Runs a command palette entry on a single click, and gives the palette a close button. Only a double-click ran a command, so clicking a row appeared to do nothing, and the palette had no way out at all for anyone using the mouse.
  • Refuses an unrecognised theme, match style, intensity, role pack or absurdity instead of quietly substituting a default. A misspelled theme used to switch to a different one and report success, and a misspelled generation option rewrote the scenario and cast while answering "generated".
  • Writes a report when the app closes unexpectedly, to the logs folder beside the sessions. Until now a crash left no dialog, no log, and nothing to send on.
  • Keeps a provider error message readable. A base URL pointing at something that is not a provider replies with a web page, and the whole of it used to arrive in the status line.
  • Lets the PowerShell control plane press keys and type into fields without needing the window in front, and adds palette listing and run commands. Keyboard shortcuts that open a dialog, including F1 and Ctrl+K, no longer leave a command waiting for someone to dismiss it.
  • Keeps the WPF PowerShell plane enabled by default, retains the self-screenshot command and bundled SearXNG Internet backend, and remains a self-contained Windows installation with no separate .NET Desktop Runtime requirement.

Assets

  • AI Arena Setup 0.4.122-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.121-beta

AI Arena 0.4.121-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 28 Jul 01:59
1239255

AI Arena 0.4.121-beta

Highlights

  • Adds a command palette on Ctrl+K. It lists what the shell can do with the shortcut shown beside anything that has one, searches by intent as well as by name so "colour" finds the themes and "find" finds transcript search, and hides commands that cannot run right now rather than offering them greyed out. Every entry runs the same action as the matching button, so the two cannot drift apart.
  • Reports shell changes to the PowerShell control plane whichever way they happen. Watching the live event stream previously only showed what an operator did through PowerShell: opening Match Setup with F2, toggling the right rail, or changing theme from the picker announced nothing at all. Navigation, overlay, rail, theme and view-preset changes are now reported for keyboard and mouse too, and each change is still reported once.
  • Stops showing normal states in the colour reserved for problems. A first run with no saved sessions was greeted by a red warning before anything had happened, and turning the Internet off reported a fault for a setting that had just been chosen on purpose.
  • Keeps the WPF PowerShell plane enabled by default, retains the self-screenshot command and bundled SearXNG Internet backend, and remains a self-contained Windows installation with no separate .NET Desktop Runtime requirement.

Assets

  • AI Arena Setup 0.4.121-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.120-beta

AI Arena 0.4.120-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 27 Jul 21:56
a837351

AI Arena 0.4.120-beta

Highlights

  • Updates the article reader to SmartReader 0.11.1, which moves the underlying AngleSharp HTML parser to 1.5.2. Versions before 1.5.0 carry a published advisory (GHSA-pgww-w46g-26qg, medium severity) in how they handle a malformed annotation-xml element, and that parser is what reads pages fetched from the Internet, so it is worth being current.
  • Publishes the resolved dependency graph to GitHub on every change to the solution. Reading the project files alone showed only the four packages named directly; the packages those in turn pull in were invisible, which is how the AngleSharp advisory went unnoticed. Vulnerability alerts are now enabled for the repository.
  • Keeps the WPF PowerShell plane enabled by default, retains the self-screenshot command and bundled SearXNG Internet backend, and remains a self-contained Windows installation with no separate .NET Desktop Runtime requirement.

Assets

  • AI Arena Setup 0.4.120-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json

AI Arena 0.4.119-beta

AI Arena 0.4.119-beta Pre-release
Pre-release

Choose a tag to compare

@neeveew neeveew released this 27 Jul 21:35
cce0dcd

AI Arena 0.4.119-beta

Highlights

  • Adds function-key navigation: F2 opens Match Setup, F5 reloads the session from disk, F7 shows or hides the right rail, F8 opens the transcript view menu, F9 starts or pauses Auto Chat, and F10 opens App Settings. Ctrl+1, Ctrl+2, and Ctrl+3 switch between AI Lab, Agent, and AI Collaborate. Press F1 for the full list.
  • Keeps destructive actions off the keyboard: Reset remains a deliberate pointer action, and no shortcut can discard a run.
  • Hides sessions that never received a turn from the session picker, behind a "Show sessions with no turns" toggle, so a data root shared with another AI Arena build no longer buries real sessions. The current and default sessions always stay listed.
  • Speeds up the session list further by only paying for the details each caller actually needs; listing a large data root is now roughly twice as fast again.
  • Folds duplicated provider request helpers and five separate text-truncation helpers into shared ones, which also settles truncated text ending three different ways and stops a length limit being exceeded.
  • Keeps the WPF PowerShell plane enabled by default, retains the self-screenshot command and bundled SearXNG Internet backend, and remains a self-contained Windows installation with no separate .NET Desktop Runtime requirement.

Assets

  • AI Arena Setup 0.4.119-beta.exe
  • SHA256SUMS.txt
  • changelog.md
  • changes.txt
  • github-release-notes.md
  • release-checksums.sha256
  • release-manifest.txt
  • release-signing.json